EValuation process on retraining displays "old" labels from Coco

General
1

With a custom yolov5 model with 2 classes renamed as myclass0 and myclass1, the hailomz eval is providing very bad results. When I look into the generated video, I see that my objects are detected, but classified as “person” and “bycicle” (ie the old COCO labels names).
I double check my .pt and .onnx files, and they have the proper labels.
So I have the feeling that the eval script is using some previous class label, hence producing wrong AP results at the end.
So first question : how to specify new label names for the eval stage ?

2

Hey @sylvain.fabre ,

You’re right — the evaluation is pulling in the wrong class labels. By default, the Hailo Model Zoo (hailomz) uses the labels defined in the YAML config. If that file still lists the original COCO labels, you’ll end up seeing results like “person” or “bicycle” even if your model was trained on something completely different.

How to Fix It

You need to make sure your YAML config reflects that.

Update Your YAML Config

In the YAML file you’re using with hailomz compile and hailomz eval, update these fields:

1. evaluation.classes

Set this to the number of classes — for example, 2 if you have two.

2. postprocessing.classes

Same thing — set to 2.

3. postprocessing.class_names

This is the important one. Set your custom class names like this:

postprocessing:
  classes: 2
  class_names:
    - myclass0
    - myclass1

This makes sure both the post-processing and evaluation stages know what your model’s outputs mean.

Example Snippet

If you’re using a config file like yolov5m_custom.yaml, it should include:

evaluation:
  classes: 2

postprocessing:
  classes: 2
  class_names:
    - myclass0
    - myclass1

Also double-check that this file is the one you’re actually passing to the --yaml flag in your commands.

1 Like